AI Buzzword Olympics – Event #1

"We will Train Our Own Model"

The three words people use when they want to sound AI-smart. Every AI strategy meeting has one. Someone confidently says: "We will train our own model." Heads nod. The architecture diagram gets updated. The AI strategy suddenly looks very sophisticated. Nobody asks the uncomfortable question:

"What exactly do you mean by train?": Because training a model sounds impressive. Understanding what it actually involves is less glamorous. A model doesn't become intelligent because we uploaded a few documents and clicked Train. Training means taking huge amounts of data, feeding it through a model, calculating how wrong its predictions are, and adjusting millions or billions of parameters repeatedly until the model improves. That is not a weekend project.

It requires massive datasets, data preparation, GPUs, distributed computing, experimentation, evaluation, monitoring, and a lot of money. And the bill doesn't stop at the GPU. You need people who understand the data. People who understand ML. People who understand distributed systems. People who can evaluate whether the model actually improved. And people who can explain why the model suddenly became worse after spending another million dollars making it "better."

Then comes the classic enterprise statement: "We will train it on our 20 years of company data." Wonderful. Is the data clean? Relevant? Labelled? Current? Consistent? Duplicate-free? Free from confidential or sensitive information?

Do you even know what is in those 20 years of data? Having 10 million documents is not the same as having 10 million useful training examples. That is not an AI problem. That is a data engineering problem wearing an AI badge. And here is the part that gets conveniently skipped in meetings:

Training from scratch can take weeks or months, depending on the model, dataset, hardware, and training objective. And the cost? It can range from thousands of dollars for relatively small experiments to millions of dollars for large foundation-model training runs before you even account for engineering, data preparation, experimentation and failed runs.

Yes. Failed runs. Because training is NOT: Upload → Train → Congratulations.

It is more like: Prepare the Data → Train → Evaluate → Fail → Tune → Train again → Evaluate → Fail differently → Repeat.

That is engineering. So before proudly announcing : "We will train our own model."

Ask five questions:

1. What problem are we solving? 
2. What data will teach the model about the problem? 
3. How will we measure improvement? 
4. What infrastructure will we need? 
5. And what will it cost?

If those answers are not clear, you are probably not discussing an AI strategy. You are discussing AI vocabulary. And there is nothing wrong with using an existing foundation model. In fact, that is often the smarter engineering decision. Use a strong foundation model. Add RAG. Improve your data. Fine-tune when there is a real reason. Train from scratch only when there is a compelling business and technical justification. Because good engineers don't build everything themselves. They know what is worth building and what is NOT. So the next time someone says: "We'll train our own model." Don't roll your eyes.

Just ask: "Great. What exactly are we training, on what data, using how much compute, for how long and what is the budget?" That usually separates AI strategy from AI theatre.

In Conclusion, Some people have mastered the art of sounding technical without actually being technical. They can talk about AI agents, foundation models, fine-tuning, RAG, GPUs and autonomous systems for an hour. Ask them to explain the architecture, you will hear Silence. Ask about the data, Silence.

Ask about cost, scale, failure modes or operational ownership, it will lead to another strategy meeting. Buzzwords are not technical depth. Architecture diagrams are not engineering. And saying "AI" and related buzzwords 50 times does not make a strategy intelligent. Leadership is not about knowing every technology. It is about knowing enough to ask the right engineering questions and having the courage to admit when you don't know the answer. Because buzzwords can win a meeting. Engineering wins production.